Automation and Digitalization of Polymerflood Viscosity Surveillance Through Machine Learning– A Case Study from Marmul

Abstract

Abstract Since its commencement in 2010, Marmul polymer EOR has been one of the worldwide successful full field applications. One of the key success factors for the project is maintainingwellhead viscosity at the target, which has been monitored by daily selective wellhead sampling. However, daily sampling covers only 7% of the polymer injectors. Recently, a digitalization project to enhance viscosity monitoring was successfully completed. One of the outcomes is utilizing the digital data available in field to have a live viscosity of all polymer injectors using an empirical power law model along with a calibration factor. Machine learning will handle any deviation of these readings by a well-established sampling program to continually re-calibrate the model.In this paper, the approach and outcomes of this projectare shared. Two polymer injectors are selected as a demonstration of the concept and main outcomes. Statistical evaluation was used to initially select the determining process parameters such as wellhead concentration, flowrate, tubing-head pressure, and tubing-head temperature. It has been concluded that wellhead polymer concentration is highly correlated to measured wellhead viscosity. The measured viscosities in the last two years (2020 and 2021) for each well were divided into; a training set (~65%) and a test set (~35%). The training set is used to calculate the calibration factor, while the test set is used to validate model predictions. Out of 415 date points, the average viscosity of polymer injectors MMPI-1 and MMPI-2 are 20.7 and 23.1 cP, respectively. The standard deviation of the measurements of injectors MMPI-1 and MMPI-2 are 3.3 and 4.8 cP, respectively. Viscosity was correlated to wellhead concentration by a power law model with experimentally obtained constant and law's exponent. Using the training set, a tuning parameter, α, was appliedwith criteria of minimummean absolute error (MAE) for each injector. α determined of MMPI-1 and MMPI-2 is 0.915 and 0.981, respectively. The model resulted in good predictions with an average MAE of around 20%. Furthermore, the model proved to be robust and reliable to be applied for live viscosity readings of all Marmul polymer injectors. Machine learning is essential for future tuning of the model for all polymer injectors in Marmul based established program of wellhead measurements. The outcomes of this digitalization and automation step in polymerflooding has demonstrated significant, positive impact on optimization of chemicals, resources, and the overall reservoir management. This work is setting another milestone in the utilization of data analytics and digitalization of fullfield polymer EOR. Machine learning coupled with excellent metering and data streaming have shown added value to overall project management. This is more critical with the shift towards agile work environment and net zero. Significant opportunities have been already realized as an outcome of this project such as quantification of polymer overdosage, which triggered a work in progress to reduce any value-eroding polymer dosage. In Marmul, the improved surveillance of wellhead viscosity and timely optimization of polymer dosage have already positively impacted project economics, GHG and HSE.